WAN 2.2 – T2V – LightX – 4 Steps – UltraFast – ULTIMATE Workflow – AUTO PROMPT – LLM Agent – Fully Local – HD – 16 VRAM – 64 RAM – Sage Attention – Torch Compile
The WAN 2.2 workflow is the ultimate solution for text-to-video (T2V) creation. It combines cutting-edge performance with extreme ease of use, while remaining 100% local for total control and maximum privacy. It leverages the latest innovations such as LightX, optimized LoRAs, Sage Attention, Torch Compile, as well as a local LLM agent for automatic prompt generation.
🚀 Main Features:
WAN 2.2: Ultra-stable and fast T2V video generation engine.
LightX Acceleration: Dramatically reduces rendering time while maintaining exceptional visual quality.
Sage Attention: Advanced attention management for improved coherence in video outputs.
Torch Compile: Automatic performance optimization via PyTorch dynamic compilation, for an even faster and smoother workflow.
4-Step Process: Simple and quick setup.
UltraFast Rendering: Designed for high-throughput production with maximum efficiency.
Auto Prompt with LLM Agent: Automatic generation of optimized prompts through a local LLM, minimizing manual input.
Fully Local Setup: Operates entirely offline with no cloud dependency.
LightX LoRA Models: Two lightweight yet powerful LoRA models for high-quality visual generation.
High & Low Noise Models: Fine-tune quality and style with two distinct noise model types.
HD Output (1280×720p): Crisp and detailed high-definition video.
Optimized for 16 GB VRAM and 64 GB RAM: Fully leverages hardware resources for smooth operation.
💡 Workflow Advantages:
⚡ Ultra-fast rendering, even on complex scenes.
🤖 Automatically generated prompts to save time.
🔒 Zero cloud dependency: everything runs locally with full data control.
🎯 Enhanced visual coherence thanks to Sage Attention.
🔧 Advanced performance optimization with Torch Compile.
🎥 HD video quality for professional-grade results.
🖥 Recommended Setup:
GPU: Minimum 16 GB VRAM
RAM: 64 GB recommended
Video Output: 1280×720p
Description
✨ Key New Features in v1.1:
WanVideo Enhance-A-Video Low: New integrated node that automatically enhances video quality while keeping performance impact to a minimum.
GGUF Model Support: Advanced model optimization for faster and more accurate rendering.
Minor Improvements: Stability and smoothness tweaks for an even better user experience.
FAQ
Comments (12)
The height and width nodes are cross-wired such that height controls width and width controls height. Easy fix.
Yes, I haven't found how to make a switch to go from (16/9 to 9/16)
Something is screwed up with the multi-lora loader on the high side such that it was greyed out and not functioning. Reloading the node resolved the issue, but it caused quite a bit of confusion!
ok thanks for the feedback I find this very strange, I mainly use multi lora low to improve details, I find it works better like this for me. Have fun.
It's unreal how optimized this workflow is! I'm achieving higher resolutions with this workflow than I've ever achieved with any others that I've tried. Generating 6 second videos at 1280x720 in under 10 minutes on a 4090. I'm astonished at the wizardry that allows this to all fit into the VRAM I have. VERY high fidelity stuff.
I ended up ripping out the LLM stuff as it was slowing things down and I'm not sure it is better at writing prompts than I am. Consider releasing a version without the LLM or a toggle to disable it.
Thanks for your feedback, have fun!
SKroUserIA Definitely do a version without the LLM model - and perhaps something that would work with 12GB GPUs? I can run it with my RTX 5070 but it takes a very long time.
@kroms50 Not having a 12 vram gpu I will never be able to ensure that everything works as expected, and for the version without llm I am currently working on it, an update will arrive soon with some additional additions to improve the quality. I am in the testing phase currently.
@SKroUserIA Epic! thank you so much for the response! Can't wait to mess around with the model!
I will be abandoning the version using Wrapper, which has become unstable in terms of VRAM for a reason that is still unknown. An update will be coming soon with a more stable "native" version (2.0), which will offer several variants (with or without LLM). I wanted to inform you that there will therefore be no more updates for this workflow based on Wrapper. Thank you for your support and see you soon to test the new workflow.
I'm now suspicious of everything KJ does- an enthusiastic amateur who gets many things plain wrong. To get the most out of 16GB of VRAM use a NATIVE workflow: try launching ComfyUI with "--disable-smart-memory --cache-none --reserve-vram 8.4 --lowvram". You need 64GB+ of system RAM- to be honest if you do not have this much RAM at least, you should give up any idea of doing good video gen at a decent speed.
The DOWNSIDE of launching Comfy with these arguments is that Clip will happen on your CPU, adding maybe 60secs of overhead, but the iteration rate will be unchanged. And your VRAM can be maxxed out with latent data, allowing the most frames at the best rez, in linear time.
@blobby99 I just released a new workflow called native v2.0, it works much better ;) after that I do not criticize his work because any contribution of talent is appreciated. Rare are those who can develop nodes with such technical competence